[molpro-user] gmainv1 failure to allocate - what to do?

[Manhui Wang]

Dear Grigory,

On 31/01/14 19:08, Grigory Shamov wrote:
> Dear Manhui,
>
> Thank you for the reply.  We have 48GB or 96GB RAM nodes, always with 
> 47GB swap partition. The memory should be allocated by Torque 
> accordingly to user's request. If he uses more memory in MolPro than 
> requested from the batch queuing system, the job gets terminated 
> automatically ( a different kind of error to see, for getting SIGTERM, 
> and the event logged for us).
>
> The particular job ran on four 48GB nodes, four processes each (3 
> compute, 1 data server as it was SF version) and requested 98 GB of 
> RAM — which is not enough, but the job didn't reach the state of being 
> killed for using too much, but just failed to allocate, right?
Yes, it appeared you requested too much memory.
>
> SHMMAX is a per-node value for SystemV shared memory. Some old 
> versions of GA used to depend on it. I should've think that for using 
> SF it is irrelevant, unless MolPro uses also SystemV internally.
For pure MPI-2 version of parallel Molpro, SHMMAX is irrelevant.
>
> On the other hand we do not allow for virtual memory overcommit. Could 
> MolPro want to allocate a large Virtual memory, much larger than the 
> Residential  memory ?
Molpro uses c/c++ allocation functions, and the OS determines how to 
allocate the memory. However, even if the OS can allocate large virtual 
memory for Molpro, the job will probably hang up or die due to poor 
performance of virtual memory.

Best wishes,
Manhui
>
> -- 
> Grigory Shamov
> HPC Analyst, Westgrid/Compute Canada
> E2-588 EITC Building, University of Manitoba
> (204) 474-9625
>
>
>
> From: Manhui Wang <wangm9 at cardiff.ac.uk <mailto:wangm9 at cardiff.ac.uk>>
> Date: Friday, 31 January, 2014 12:03 PM
> To: Grigory Shamov <Grigory.Shamov at umanitoba.ca 
> <mailto:Grigory.Shamov at umanitoba.ca>>
> Cc: "molpro-user at molpro.net <mailto:molpro-user at molpro.net>" 
> <molpro-user at molpro.net <mailto:molpro-user at molpro.net>>
> Subject: Re: [molpro-user] gmainv1 failure to allocate - what to do?
>
> Dear Grigory,
>
> On 31/01/14 15:07, Grigory Shamov wrote:
>> Dear MolPro users,
>>
>>
>> I have compiled MolPro with Intel 12 and auto-built MVAPICH2. (I had to
>> substitute MVAPICH2-2.0b instead of the unavailable tar ball of
>> MVAPICH2-1.9 in the processs; also to decrease optimization to -O2; I
>> attach my CONFIG. I understand it picked I8 array indexes, right?).
>>
>> When our user ran it on a large task, with -m 1000m he got the following
>> error:
>>
>> Multipassing necessary in transformation. To avoid, increase memory by
>> 3561.33 Mwords.
>>   ? Error
>>   ? 2-ext paging plus 3-ext ints not yet working (kintb)!
>>   ? The problem occurs in cckint
>>
>> Then, I told him to increase the memory, but with -m 4600m it fails in a
>> different way:
>>
>>
>> gmainv1 failure to allocate 4600230002
>> gmainv1 failure to allocate 4600230002
>> gmainv1 failure to allocate 4600230002
>> gmainv1 failure to allocate 4600230002
> How much memory does the machine have? If you want to run Molpro job with
> -m 4600m *with 4 processes*, actually you are requesting 4600*1000000 
> words(1 word = 8 bytes) per process.
> The total requested memory is 137 GB (4600*1000000 * 8 * 4).
>> I've googled this problem, and have seen some answers related to SHMMAX
>> value. On our cluster we have it increase though:
>>
>> cat /proc/sys/kernel/shmmax
>> 68719476736
> This is around 64GB.
>
> Could you please check if the requested memory exceeds the hard memory 
> limit on the machine?
>
>
> Best wishes,
> Manhui
>> Could you please suggest, how does one avoid this problem,? Also, I did
>> use the default limits when compiling (number of atoms etc.) -- if that is
>> the reason, which limits should be increased? Thank you very much!
>>
>>
>>
>>
>> _______________________________________________
>> Molpro-user mailing list
>> Molpro-user at molpro.nethttp://www.molpro.net/mailman/listinfo/molpro-user
>
> -- 
> -----------
> Manhui  Wang
> School of Chemistry, Cardiff University,
> Main Building, Park Place,
> Cardiff CF10 3AT, UK

-- 
-----------
Manhui  Wang
School of Chemistry, Cardiff University,
Main Building, Park Place,
Cardiff CF10 3AT, UK

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.molpro.net/pipermail/molpro-user/attachments/20140202/3b4bc6c8/attachment.html>